A Case for Rigorous Workload Classification

Traditional workload labels such as "archival" and "HPC" are poorly understood and inconsistently applied. As usage of systems has evolved, the language to describe this usage has stagnated. To better understand how workload type translates into system design requirements, we use a combination of longitudinal analysis and statistical feature extraction to categorize workload traces and study how the properties of classical workload types, such as the "write-once, read-maybe" assumption for archives, have evolved over time. Once this step is complete, we intend to move to active classification of workloads to replace these broad, poorly specified categories with quantitative metrics that can be used to improve metrics such as power, availability, and performance by mathematically relating storage algorithms with workload properties.